AI footage pipeline for direct-response ads

Ancient Remedies (Bublishing) produced direct-response ads at volume for natural-health offers in English and Spanish. The Spanish line needed presenter footage, product shots and medical animations that didn't exist as stock, fast enough to ship dozens of variations a month.

Early generated footage looked artificial, which hurt ad realism and burned generation credits on takes nobody could use. Spanish prompts drifted into English-market visuals when the starting frames and prompt framing were North American.

• A prompt system with two rule sets, UGC (handheld realism, native Spanish speakers, LATAM settings) and medical (clean animation, anatomically plausible, no text artefacts), applied to every shot before generation.
• Generation on Higgsfield / Seedance 2.0 at 3:4, 8 s, 720p, with a fallback route when a shot wouldn't resolve; Google Flow for Spanish-language dialogue shots after diagnosing the language-drift problem; ComfyUI on RunPod for experimental looks.
• The whole flow driven from Claude: the concept's shot list goes in, cleaned prompts come out, clips are generated, downloaded, audio-stripped and filed into the concept folder, numbered and ready for the edit.

• 40+ direct-response ads and 227 ad combinations in under five months, at 3.43% CTR and positive contribution margin, on $14k+ of monthly spend.
• An AI hook-ideation skill adopted by the ~8-person editing team, plus Claude-based tools that build AI footage shot lists from live performance data and fix captions automatically.
• Documented the full UGC and medical prompt system and shared it in the editor channel; the Spanish team adopted it. Coached a teammate on the workflow and ran QC on her output.
• Proposed moving the team's generation plan to annual billing with a capped auto-refill: about $670 a year saved and no more running out of credits mid-session.

Higgsfield / Seedance 2.0, Google Flow, ComfyUI + RunPod, Claude, DaVinci Resolve, After Effects.